Unsupervised anomaly detection in multi-class datasets using Generative Adversarial Networks

"Presented in this thesis is a novel Generative Adversarial Network, or GAN, based method, D-AnoGAN, for detecting anomalies in complex datasets containing disconnected data manifolds. Current state-of-the-art methods treat disconnected data manifolds as a single, continuous one to learn from. The key contribution of D-AnoGAN is specifically accounting for the discontinuity between manifolds within a dataset during training. To achieve this, a multi-generator network is first implemented, where each generator is responsible for learning a unique manifold of data. Second, a machine learning mechanism called a ''bandit"" is implemented to find the optimal set of generators required to cover all data manifolds through unsupervised prior-learning. Finally, the multi-generator and bandit are used to cluster data from the same manifold together during training, allowing them to be learned in a disconnected fashion. The proposed method's effectiveness is demonstrated on two publicly available datasets, as well as a new experimental dataset developed in-house, where state-of-the-art results are achieved."

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Aerospace Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2021

Author dc:creator

• Dimon, Walker Lee

Contributors dc:contributor

• Lembeck, Michael F
• Tran, Huy T

• Generative Adversarial Networks, Unsupervised Anomaly Detection, Computer Vision, Unsupervised Learning, Machine Learning

Statement dc:rights

• Copyright 2021 Walker Dimon

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/110594

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/110594